Feature Request: Independent model assignment for librarian and fixer (break the inheritance chain)

[BetterToAutomateTheWorld]

Feature: Independent model assignment for librarian and fixer (break the inheritance chain)

Problem

Currently, when fixer has no explicit model set, oh-my-opencode-slim inherits the librarian's model. This makes it impossible to have:

• librarian → cheap local model (e.g. ollama/qwen3.8:27B) for research/doc lookups
• fixer → frontier/session model (e.g. zai-coding-plan/glm-5.2) for implementation work

Setting both explicitly works, but there's no way to say "use the session/orchestrator model" for one agent while pinning another to a local cheap model. The only workaround is an external plugin wrapper that strips the inherited model after the fact.

I want/need this librarian/fixer decorrelation because my cheap model is ollama local model, such as ollama/qwen3.8:27B, and i noticed too many issues when qwen is the fixer, I prefer to avoid it, but it's a really good model for other purpose (librarian, git-ops tasks, tracer, etc.). Maybe someday qwen 4.x or 5.x could fixes these issues in still a small distilled version, finger crossed :).

Current behavior (the inheritance chain)

librarian.model = "ollama/qwen3.8:27B"
fixer.model    = (not set)
  → fixer silently inherits librarian.model  ← unwanted coupling

Desired behavior

Option A — explicit sentinel (preferred):

{
  "agents": {
    "librarian": {
      "model": "ollama/qwen3.8:27B"
    },
    "fixer": {
      "model": "$session"   // or "$primary", "$orchestrator"
    }
  }
}

$session / $primary / $orchestrator would mean "follow whatever model the current session/orchestrator is using", breaking the librarian→fixer inheritance.

Option B — config flag (simpler):

{
  "agents": {
    "librarian": {
      "model": "ollama/qwen3.8:27B"
    },
    "fixer": {
      "model": "ollama/qwen3.8:27B",
      "inheritModelFrom": "session"   // or "orchestrator", "primary"
    }
  }
}

Option C — convention (zero-config):

If fixer has no model and librarian has an explicit model, treat fixer as "use the session model" rather than inheriting from librarian. This is the most intuitive behavior: the implementation agent should default to the session's main model, not the research agent's cheap local model.

Why this matters

• Cost control: Librarian does heavy web/doc research on a cheap local model; fixer does code implementation on a frontier model. Mixing them means either expensive research or weak implementation.
• Separation of concerns: Research quality and coding quality have different model requirements. Forcing them to share a model couples two unrelated cost/quality trade-offs.
• Current workaround is fragile: Requires a custom plugin (oh-my-opencode-slim-plugin.ts) that intercepts the config hook and strips fixer.model post-build. This breaks on any internal refactor of the agent config shape.

Workaround today (for reference)

// oh-my-opencode-slim-plugin.ts
const wrapper = async (input: any) => {
  const hooks = await slim(input)
  if (Array.isArray(hooks.agent)) {
    hooks.agent.forEach(a => {
      if (a.name === "fixer" && a.config) a.config.model = undefined
    })
  }
  // also strip after config hook runs
  const originalConfig = hooks.config
  if (typeof originalConfig === "function") {
    hooks.config = async (cfg: any, ...rest: any[]) => {
      const result = await originalConfig(cfg, ...rest)
      if (cfg?.agent?.fixer) cfg.agent.fixer.model = undefined
      return result
    }
  }
  return hooks
}

Related

• The inheritance behavior is documented implicitly in the codebase but not as a user-configurable feature.
• Similar use case: any user who wants a cheap research lane + a strong coding lane in the same session.